The MetaFilter discussion of Google’s people-sized LOGO turtle offers up plenty of scenarios for where and how self-driving cars might be most useful once they get outside their Mountain View test track:
- a way to reduce the horrific number of deaths and injuries on the roads
- a way to make car-sharing the norm and reduce congestion
- a replacement for taxi drivers 
- a replacement for long-haul truck drivers
- a replacement for paratransit services
- a source of liberation for those currently confined by sprawl and inadequate transport options: the young, the old, the poor, the disabled
- a last-mile connection to public transit networks
- a replacement for public transit networks, leaving the last mile for individual drivers
- a reinvention of urban transportation
- a reinvention of rural transportation
- and so on
All well and good, but we’re not going to wake up one morning to find all our transportation devices turned sentient and electric. Horses were still used for road haulage decades after carriages went horseless. So right now there’s a lot of ‘better horse’ thinking, often at cross-purposes, when you’d assume that the future will, as always, be unevenly distributed. Completely autonomous vehicles will come to small, tightly-controlled environments first and the second-order effects of those deployments will shape subsequent use, by which time partial implementations that trickle down to existing vehicles will have already changed the conditions under which we might want them to take full control.
In broader terms: a potentially disruptive technology has to succeed at doing some of the things before it can set its sights on all of the things, and those first implementations and instantiations lay a foundation for what follows, not just by what they do but in what they choose not to do. They establish focus. That we can come up with so many different applications right now may actually be an impediment compared to technologies that emerge with simple goals and expand through the discovery of further applications and iterative enhancement to accommodate them.
The gap between minimum viable prototype and maximum imaginable future is a hard one to cross.
 Perhaps with the aid of a Boston Dynamics bulldozer that will intelligently render cities more like Mountain View.
 That will need to address the Bodily Fluids Problem, i.e. when someone summons an Ubermatic to take them home after a night out and they throw up en route. Or worse.
 SF writers, you can have that one for free.
 Big Amazon-style warehouses, obviously, but I can also imagine airports rapidly deploying self-driving vehicles to shuttle passengers, luggage and equipment across the tarmac: it’s another private space with highly schematised routes driven at low speed, and it’s a good way to get the public comfortable with the tech in a controlled environment. The Dulles mobile lounge, an artefact of the future-in-the-past, may well have a second life.
 Your average eighteen-wheeler is now a kind of Taylorised cyborg, thanks to GPS navigation and tracking and telemetry and anticipatory gear selection and traffic monitoring and a couple of decades of logistics work that calls for specific routes hitting specific waypoints at specific times; we’re long past the days of the CB-era ‘open road’ where the bosses didn’t care too much about the journey as long as you reach your destination on time.
 For some reason I think of Ted Nelson’s Xanadu here, versus the CERN or Mosaic-era web.